Cost-Based Optimization of Integration Flows - Datenbanken ...

5 Multi-Flow Optimization
example partition tree of height h = 2. On the first index level, the messages are partitioned
according to customer names ba 1 (m i ), and on the second level, each partition is divided
according to the range of order total prices ba 2 (m i ). Horizontal partitioning with the
temporal order of partitions reason the outrun of single messages. However, the messages
within an individual partition are still sorted according to their incoming order.
Based on the partition tree data structure, there are two fundamental maintenance
procedures of such a horizontally partitioned message queue:
• The enqueue operation is invoked by the inbound adapters whenever a message was
received and transformed into the internal representation. During message transformation,
partitioning attributes are extracted with low cost as well. Algorithm 5.1
describes the enqueue operation. We use a thread monitor approach for synchronization
of those enqueue and dequeue operations in order to avoid any busy waiting.
Subsequently, if a partition with ba l (m i ) = ba(b i ) already exists, the message is inserted;
otherwise, a new partition is created and added at the end of the list. The
recursive insert operation depends on the partition type, where we distinguish node
partitions (index levels 1 to h − 1) and leaf partitions (last index level that contains
the messages). In case of a node partition, the algorithm part line 3 to line 13 is
used recursively, while in case of a leaf partition, the message is simply added to the
end of the list of messages in this partition.
• The process engine then periodically invokes the dequeue operation according to the
computed waiting time ∆tw. This dequeue operation removes and returns the first
top-level partition with b − ← b 1 . The removed partition exhibits the property of
being the oldest partition within the partition tree with t c (b 1 ) = min |b|
i=1 t c(b i ), which
ensures that starvation of messages is impossible.
The operations enqueue and dequeue are invoked for each individual message. Therefore,
it is meaningful to discuss the worst-case complexity for both operations. Let sel(ba i )
Algorithm 5.1 Partition Tree Enqueue (A-PTE)
Require: message m i , partitioning attribute values ba(m i ), index level l
initialized counter c ← 0, boolean SEB (serialized external behavior)
1: while ∑ |b|
j=1 |b j| > q do // ensure maximum queue constraint q
2: wait // wait for notifying dequeue() (without busy waiting)
3: for j ← |b| to 1 do // for each partition
4: if ba l (m i ) = ba(b j ) then // if partition already exists
5: b + ← b j
6: break
7: else if SEB then
8: c ← c + |b j | // count outrun messages for later serialization
9: if b + = NULL then // if partition does not exist
10: b + ← create partition with t c (b + ) ← t i (m i )
11: b |b|+1 ← b + // add as last partition
12: m i .c ← c // zero in case of newly created partitions or for disabled SEB
13: b + .insert(m i , ba(m i ), l − 1) // recursive insert
14: notify // notify waiting dequeue()
136